[
https://issues.apache.org/jira/browse/FLINK-15447?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17006558#comment-17006558
]
Victor Wong commented on FLINK-15447:
-------------------------------------
[~yunta] [~klion26] Thanks for your reply :)
_What caused the "/tmp" directory full?_
---
It may not be caused by Flink applications since other applications like
MapReduce/Spark would run on the same Yarn cluster too.
I think it's not safe to use "/tmp" directory since it's shared by other tasks.
Tasks, like the Flink TaskManager, could make use of its own working directory
assigned by Yarn.
_Flink already set io.tmp.dirs as 'LOCAL_DIRS' in YARN_
---
This config can not solve the problem completely, since the third-party
libraries used by users may still use "java.io.tmpdir" property.
There is similar discussion here:
https://issues.apache.org/jira/browse/HADOOP-2735
> Change "java.io.tmpdir" of JM/TM on Yarn to "{{PWD}}/tmp"
> -----------------------------------------------------------
>
> Key: FLINK-15447
> URL: https://issues.apache.org/jira/browse/FLINK-15447
> Project: Flink
> Issue Type: Improvement
> Components: Deployment / YARN
> Affects Versions: 1.9.1
> Reporter: Victor Wong
> Priority: Major
>
> Currently, when running Flink on Yarn, the "java.io.tmpdir" property is set
> to the default value, which is "/tmp".
>
> Sometimes we ran into exceptions caused by a full "/tmp" directory, which
> would not be cleaned automatically after applications finished.
> I think we can set "java.io.tmpdir" to "{{{{PWD}}}}/tmp" directory, or
> something similar. "{{{{PWD}}}}" will be replaced with the true working
> directory of JM/TM by Yarn, which will be cleaned automatically.
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
